KI kann deine Belege lesen,
sie kann sie nur nicht behalten

Forward Claude a receipt and it reads it instantly. It knows the merchant, the amount, the currency, even which category the purchase belongs in. The parsing is genuinely impressive. Then close the chat, come back next week, and ask what you spent. It has no idea. The receipt, the amount, the whole exchange is gone.

The thing missing here isn't intelligence. It's memory. An AI assistant can understand your money in the moment and retain none of it afterward. That gap is exactly what a finance tracker is for, and it's a useful lens for thinking about what these tools actually do.

Reasoning is not memory

Large language models are stateless by default. Each conversation runs inside a context window, and once that window ends, the model keeps nothing. The next chat starts from zero. This is why the same assistant that just categorized your lunch can't tell you your monthly food total: it never stored the lunch anywhere it can read later.

Newer "memory" features in chat assistants help a little, but they're built for preferences and facts, things like "I live in Dubai" or "I prefer metric units." They are not a transaction ledger. Remembering that you spent 47 AED on groceries on June 2nd, alongside two hundred other entries you can total and filter, is a different problem. It needs a ledger, not a note the model occasionally recalls.

What "financial memory" actually means

Financial memory is a durable, structured record of every transaction: the amount, the original currency, the category, the description, the date, and the exchange rate at the time it happened. Structure is the important word. Storage alone isn't enough.

Pasting your expenses into a notes app is storage. But ask "how much did I spend on food in March across three currencies" and a wall of text can't answer. The data has to be parsed into fields a computer can sum, convert, and filter. Without that, every question forces the AI to re-read and re-add a pile of text, and the numbers drift each time it tries.

Claude reasons. Kachink remembers.

This is the division of labor Kachink is built around. Claude is the reasoning layer: it reads what you type, parses a receipt, figures out the currency and category, and understands the question you're asking. Kachink is the memory layer: it takes that understanding and stores it as a real transaction in a real ledger, so the record survives long after the conversation does.

Kachink runs as a remote MCP server, the open standard Anthropic introduced for connecting AI assistants to external tools. When you say "spent 47 on groceries," Claude calls Kachink, which writes the entry, converts it to your base currency at that day's rate, and keeps it. Ask about any month later, even months later, and you get the same numbers every time, because they were stored, not re-guessed.

Why this is the whole point

Once you separate reasoning from memory, the value of each part gets clearer. The AI's job is to make logging effortless: no forms, no dropdowns, just plain language inside a chat you're already in. The tracker's job is to never forget and never miscount. You don't want your assistant to be clever about your finances. You want it to be correct, and correctness over time requires memory.

It also points at where this goes next. A receipt in your email or a payment text on your phone is just unparsed data. Claude can read either one and hand the structured result to Kachink in a single step, turning a message you'd otherwise ignore into a stored, queryable transaction. The reading is the AI's strength. The remembering is the part you've been missing.

If you've been impressed by how well Claude understands your spending but frustrated that it forgets, that's the gap to close. Give it a memory and the rest follows. See how logging works, or read why bank sync isn't the answer either.